[Tinyos-devel] Re: CTP - day and night

[Omprakash Gnawali]

On Jan 5, 2008 6:06 PM, Omprakash Gnawali <gnawali at usc.edu> wrote:
>
> On Jan 5, 2008 4:32 PM, Philip Levis <pal at cs.stanford.edu> wrote:
> >
> > On Jan 3, 2008, at 9:16 PM, Omprakash Gnawali wrote:
> >
> > > I ran CTP for 24 hrs - 4pm till 4pm the following day. The result
> > > tagged "aaa9" has been added to this page:
> > > http://enl.usc.edu/~om_p/net2/ctp4bitle/figures.html
> > >
> > > In order to put results from experiments done in the past on the same
> > > graphs, I have changed the time scale for the past results. You can
> > > read this on the key for the lines on the graph. For example, time * 2
> > > means time scaled by 2. The results seem qualitatively similar to aa9,
> > > which was the previous long experiment. The numbers are on the top of
> > > the page and the graphs follow. I think it would be interesting to
> > > understand the cause for loops with aaa9. It would be also interesting
> > > to understand the temporal variation in various metrics, including
> > > loops. Having data from a 24-hr experiment allows us to study such
> > > temporal variations.=
> >
> > Just to check -- aa9 has the < rule for loop detection, while aaa9
> > has the <= rule? There's a big difference in delivery ratio. Also,
> > aaa9 has a good number of queue full events... it would be good to
> > look into what's causing them. Is it because of how CTP pauses on
> > loop detection?
> >
> > Phil
> >
>
> Yes, aaa9 uses the <= test while earlier experiments used the < test.
> One of the reasons why the results look more dramatic that they really
> are is because of the difference in time scale used in the graphs for
> lines corresponding to aa9 vs aaa9.
>
> I just added a figure called cumfullevents(t) which is below the loop
> graph to see if there are more full events when there are more loops.
> The hypothesis is not supported beyond T=30k. After this point, the
> full events taper off but not the loop detections. Something special
> happened in the first 8 hrs. I need to look at temporal variation in
> delivery ratio - my guess is we are seeing a significantly lower
> delivery ratio in the first 8 hrs which explains the 2% difference in
> mean delivery ratio over 24 hrs.
>
> Here is the link again:
> http://enl.usc.edu/~om_p/net2/ctp4bitle/figures.html
>
> - om_p
>


To try to explain the lower delivery ratio with aaa9 vs aa9, I plotted
the evolution of delivery ratio. It is the last graph on this page:
http://enl.usc.edu/~om_p/net2/ctp4bitle/figures.html

I computed delivery ratio statistics on a window size of 3600s,
sliding the window by 1000s - so that's 85 data points for each node.
The overall delivery ratio is lower than that for aa9 (same setting,
different day) is due to what happened in the first 30 K seconds of
this run.

- om_p